{"id":"efaed787-fc52-4000-99a5-21aa83a7844c","arxiv_id":"2606.23079","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AdaReP is a training-free wrapper that dynamically adjusts replanning tolerance in neural MPC using mismatch deviation and sensitivity to reduce planner queries while preserving task performance.","lead":"AdaReP adapts the replanning frequency in neural world-model MPC by estimating current deviation from a cached plan and local dynamics sensitivity. A smart generalist might read it for ways to cut computation in learned-model robot control without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Perturbation-based dynamic-regret framework's capture of stale-plan penalties under neural model mismatch remains the unverified foundation for the adaptation rule.","rationale":"The reader's weakest assumption directly identifies the load-bearing theoretical step that justifies the practical wrapper; the empirical results alone do not close the gap without confirming the derivation.","tokens_in":1676,"tokens_out":257,"duration_ms":15907,"concrete_test":"Extract the dynamic-regret bound and perturbation expansion from the analysis section; independently re-derive the online adaptation rule for replanning tolerance and check whether it matches the exact rule implemented in AdaReP (including how the local sensitivity estimate is obtained).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the analysis showing that stale-plan penalties scale with reuse tolerance, accumulated mismatch, and local dynamics sensitivity, from which AdaReP's training-free online rule is derived. If the perturbation analysis assumes local linearity or bounded higher-order terms that fail for typical neural world-model rollouts (especially in image or latent space), the derived tolerance adaptation may not systematically reduce queries while preserving performance; the 80% reduction in the robot study could then be driven by domain-specific tuning rather than the framework.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes AdaReP, a training-free wrapper around neural world-model MPC that adapts the replanning tolerance online. It derives the adaptation rule from a perturbation-based dynamic-regret analysis showing that stale-plan penalties scale with reuse tolerance, accumulated mismatch since the last replan, and local dynamics sensitivity. The method requires no changes to the learned model or planner. Experiments across image-space planning, latent-space control, and a 50-trial real-world robotic manipulation study report substantial reductions in planner queries (over 80% fewer in the robot study) while maintaining comparable task performance.","tokens_in":1778,"tokens_out":392,"duration_ms":14987,"significance":"If the perturbation analysis is valid and the derived rule systematically reduces queries without performance loss, the approach would offer a practical, model-agnostic way to lower the computational cost of frequent replanning in learned-model MPC for robotics and control, without requiring retraining or architectural changes.","major_comments":[{"comment":"The central claim rests on the perturbation-based dynamic-regret framework correctly capturing stale-plan penalties under neural model mismatch (abstract and analysis sections). The provided abstract supplies no derivation details, experimental controls, statistical tests, or error analysis, so it is impossible to judge whether the data support the performance claims or whether the adaptation rule is robust when local linearity or bounded higher-order terms fail for typical neural world-model rollouts.","section":"Abstract / analysis"},{"comment":"The 80% query reduction on the 50-trial physical robot study is presented as evidence that the online rule remains effective without model changes. Without reported variance, baseline comparisons, or ablation of the sensitivity estimate, it is unclear whether the result is driven by the framework or domain-specific tuning.","section":"Experiments (robot study)"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. We address each major comment below with clarifications drawn from the full paper and note planned revisions to enhance clarity and completeness.","responses":[{"response":"The perturbation-based dynamic-regret analysis, including the derivation showing that stale-plan penalties scale with reuse tolerance, accumulated mismatch since the last replan, and local dynamics sensitivity, is fully detailed in Section 3 of the manuscript. The abstract is length-constrained and summarizes the result; we will revise it to briefly reference the analysis and its assumptions. Experiments in Sections 4–5 include controls across image-space, latent-space, and real-robot settings that empirically support the claims under realistic mismatch. We will add a discussion of the local-linearity assumption and potential robustness issues when higher-order terms dominate.","revision_made":"partial","referee_comment":"[Abstract / analysis] The central claim rests on the perturbation-based dynamic-regret framework correctly capturing stale-plan penalties under neural model mismatch (abstract and analysis sections). The provided abstract supplies no derivation details, experimental controls, statistical tests, or error analysis, so it is impossible to judge whether the data support the performance claims or whether the adaptation rule is robust when local linearity or bounded higher-order terms fail for typical neural world-model rollouts."},{"response":"We agree that variance reporting, baseline comparisons, and an ablation of the sensitivity estimate would strengthen the robot-study presentation. In the revision we will add standard deviations across the 50 trials, fixed-tolerance baseline comparisons, and an ablation isolating the sensitivity term to demonstrate that the query reductions arise from the adaptive rule.","revision_made":"yes","referee_comment":"[Experiments (robot study)] The 80% query reduction on the 50-trial physical robot study is presented as evidence that the online rule remains effective without model changes. Without reported variance, baseline comparisons, or ablation of the sensitivity estimate, it is unclear whether the result is driven by the framework or domain-specific tuning."}],"tokens_in":1326,"tokens_out":437,"duration_ms":19086,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that AdaReP derives an online rule for adjusting replanning tolerance from a perturbation-based dynamic-regret analysis and shows it cuts planner queries substantially on hardware without retraining the world model or planner. The rule uses observed deviation from the cached rollout plus a local sensitivity estimate.\n\nThe paper does a solid job on the empirical side. It reports results across image-space planning, latent-space control, and a 50-trial physical robotic manipulation study with over 80% fewer queries while holding task performance steady. Real hardware data like that is the useful part for anyone trying to run these systems on robots where constant replanning is expensive.\n\nThe softer spot is the analysis that justifies the rule. The framework claims stale-plan penalties scale with reuse tolerance, accumulated mismatch, and local dynamics sensitivity, but if the perturbation assumptions (local linearity or bounded higher-order effects) do not hold for typical neural rollouts in image or latent space, the adaptation may not be as reliable or general as stated. The abstract gives no derivation steps or error bounds, so the link between the math and the observed savings could be weaker than it appears; the robot gains might partly reflect domain-specific choices rather than the framework alone.\n\nThis is aimed at control and robotics people who already use neural world models for MPC and want lower compute on deployment. A reader working on efficiency tweaks would get concrete value from the experiments. It has enough hardware evidence and a distinct method to deserve peer review, even if the theory section requires close reading to confirm the regret scaling holds up.","headline":"AdaReP gives a training-free way to adapt replan tolerance in neural MPC using a regret bound, with clear query savings on a physical robot, though the perturbation analysis assumptions need checking.","tokens_in":2255,"tokens_out":397,"would_cite":false,"duration_ms":12639,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"AdaReP adapts replanning tolerance online from current deviation and local sensitivity to reduce MPC computation while preserving task performance.","keywords":["adaptive replanning","model predictive control","neural world models","dynamic regret","model mismatch","robot manipulation","plan reuse"],"falsifier":"In the physical robot manipulation trials, if the adaptive tolerance produces no reduction in planner queries or produces more task failures than a fixed replanning schedule, the scaling relation and derived rule do not hold in practice.","tokens_in":2589,"feed_emoji":"🤖","tokens_out":727,"duration_ms":10844,"temperature":0.7,"pith_summary":"Neural world models paired with MPC normally replan at every step to keep prediction error from growing, but this creates heavy computational cost. The paper derives how the penalty for reusing an old plan grows with reuse tolerance, accumulated model mismatch, and local dynamics sensitivity through a perturbation-based dynamic-regret analysis. From this scaling, it builds AdaReP, a wrapper that adjusts the replan threshold on the fly using only the observed deviation from the cached rollout and an estimate of sensitivity. No retraining of the world model or planner is required. Experiments across image planning, latent control, and physical robot manipulation show large drops in planner queries with little change in success rates, including more than 80 percent fewer queries in a 50-trial real-robot study.","feed_headline":"Adaptive replanning cuts MPC queries over 80% in robot tests","feed_subtitle":"Online tolerance adjustment from deviation and sensitivity keeps task performance while avoiding full replans at every step.","key_machinery":"Perturbation-based dynamic-regret framework that quantifies how stale-plan penalties scale with reuse tolerance, accumulated mismatch since last replan, and local dynamics sensitivity.","core_discovery":"We analyze the trade-off between plan reuse and prediction mismatch with a perturbation-based dynamic-regret framework and show that stale-plan penalties scale with the reuse tolerance, the accumulated mismatch since the last replanning step, and the local dynamics sensitivity. Based on this structure, we propose AdaReP, a training-free wrapper that adapts the replanning tolerance online using the current deviation from the cached rollout and a local sensitivity estimate, without modifying the learned world model or planner.","pith_inferences":["The same scaling structure could be used to set reuse thresholds in other model-based controllers that face accumulating prediction error.","Because the method needs no extra training, it may be directly applicable to already-deployed neural MPC systems on resource-limited hardware.","If the local sensitivity estimate becomes unreliable in highly nonlinear or changing environments, the adaptation may need an additional safeguard such as a maximum reuse limit."],"forward_implications":["Planner-side computation drops substantially while task success rates remain comparable across image-space planning, latent-space control, and real-world robotic manipulation.","More than 80 percent fewer planner queries occur in a 50-trial physical robot study with no loss in performance.","The wrapper works without any changes to the learned world model or the underlying planner.","The adaptation rule uses only observable deviation from the cached rollout and a local sensitivity estimate computed from the current model."],"fun_headline_variants":["AdaReP adapts replanning tolerance using mismatch deviation","Dynamic regret framework for plan reuse decisions","AdaReP wrapper adapts without changing world model","80 percent fewer planner queries in robot manipulation"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The perturbation-based dynamic-regret framework correctly captures how stale-plan penalties scale with reuse tolerance, accumulated mismatch since last replan, and local dynamics sensitivity so that an online adaptation rule remains effective.","fun_headline_variants_meta":{"raw":{"variants":["AdaReP adapts replanning tolerance using mismatch deviation","Dynamic regret framework for plan reuse decisions","AdaReP wrapper adapts without changing world model","80 percent fewer planner queries in robot manipulation"]},"model":"grok-4.3","cost_usd":0.008679,"raw_usage":{"total_tokens":3813,"prompt_tokens":629,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":86790500,"prompt_tokens_details":{"text_tokens":629,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3129,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":629,"tokens_out":55,"duration_ms":12424,"temperature":1.0,"reasoning_tokens":3129,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T08:43:59.605219+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"In the physical robot manipulation trials, if the adaptive tolerance produces no reduction in planner queries or produces more task failures than a fixed replanning schedule, the scaling relation and derived rule do not hold in practice.","supporting_citations":[],"review_version":1}